Omicverse Single Cell Cellmatch Ontology

OmicVerse Single-Cell — Cell Ontology Mapping (CellMatch)

Goal

Take an annotated single-cell AnnData whose obs[<cell_name_col>] carries free-text cell-type labels (often author-specific shorthand) and map every label to a canonical Cell Ontology (CL) term via NLP. Output is adata.obs['cell_ontology'] (matched CL term name), adata.obs['cell_ontology_cl_id'] (e.g. CL:0000084), and adata.obs['cell_ontology_score'] (cosine similarity).

Three escalating modes:

1. Plain mapping — sentence-transformer cosine-similarity match. Fastest; works on clean labels.
2. LLM abbreviation expansion — calls an LLM to expand 'TIL-1' → 'tissue-resident memory CD8+ T cell' before matching. Required when labels are short / acronym-heavy.
3. Cell Taxonomy enhancement — adds species + tissue context via the Jin 2023 Cell Taxonomy resource; matches against species-specific taxonomies as well as CL.

This skill addresses the gap noted in coverage analysis: single-cell-annotation skill handles CellTypist / SCSA / gpt4celltype annotators, but doesn't cover the post-annotation problem of mapping author-text to a standard ontology.

Quick Workflow

Mode 1 — Plain mapping (cleanest workflow when labels are full text)

1. Download the Cell Ontology to a local directory: ov.single.download_cl(output_dir='new_ontology', filename='cl.json'). ~10 MB; cached after first call.
2. Construct: mapper = ov.single.CellOntologyMapper(cl_obo_file='new_ontology/cl.json', embeddings_path='new_ontology/ontology_embeddings.pkl', local_model_dir='./my_models'). The first call computes 3000-term embeddings (30 s); subsequent calls load from embeddings_path.
3. Map: mapping_results = mapper.map_adata(adata, cell_name_col='cell_label'). Writes obs['cell_ontology'], obs['cell_ontology_cl_id'], obs['cell_ontology_score'].
4. Summary: mapper.print_mapping_summary(mapping_results, top_n=15) shows top-mapped labels and their match scores.
5. Visualise: ov.pl.embedding(adata, basis='X_umap', color=['cell_label', 'cell_ontology'], wspace=0.55, ncols=2) — side-by-side original and CL-mapped labels.

Mode 2 — Plain mapping + LLM abbreviation expansion

6. Configure LLM access: mapper.setup_llm_expansion(api_type='openai', api_key='sk-...', model='gpt-4o-2024-11-20', tissue_context='gut', species='mouse', study_context='Epithelial cells from small intestine and organoids of mice...'). The contexts inform the LLM's expansion choices.
7. Map with expansion: mapping_results = mapper.map_adata_with_expansion(adata=adata, cell_name_col='cell_label', threshold=0.5, expand_abbreviations=True). The LLM resolves 'TA-Early' → 'transit amplifying cell, early' first, then maps that to CL.
8. Use api_type='custom_openai' with base_url=... for OpenAI-compatible endpoints (Azure, Ollama, OhMyGPT, etc.).
9. Cache: abbreviation expansions are cached locally; subsequent runs reuse the cache (controlled by cache_file in setup_llm_expansion).

Mode 3 — Cell Taxonomy enhancement (species + tissue aware)

10. Load the Cell Taxonomy resource (Jin 2023): mapper.load_cell_taxonomy_resource('new_ontology/Cell_Taxonomy_resource.txt', species_filter=['Homo sapiens', 'Mus musculus']). ~30k entries; species filter avoids loading non-relevant taxa.
11. Map with taxonomy: enhanced_results = mapper.map_adata_with_taxonomy(adata, cell_name_col='cell_label', new_col_name='enhanced_cell_ontology', expand_abbreviations=True, use_taxonomy=True, species='Mus musculus', tissue_context='Gut', threshold=0.3). Combines CL match + Cell Taxonomy match.
12. Visualise: ov.pl.embedding(adata, color=['cell_label', 'cell_ontology', 'enhanced_cell_ontology', 'enhanced_cell_ontology_taxonomy_match', 'enhanced_cell_ontology_ct_id'], ...) — full provenance.

Interface Summary

ov.single.download_cl(
    output_dir: str = 'new_ontology',
    filename: str = 'cl.json',
)
# Downloads CL to <output_dir>/<filename>; idempotent.

ov.single.CellOntologyMapper(
    cl_obo_file: str | None = None,
    embeddings_path: str | None = None,
    model_name: str = 'all-mpnet-base-v2',     # sentence-transformer
    local_model_dir: str | None = None,         # cache dir for the ST model
    auto_download: bool = True,
)

Core methods (all use cosine similarity at threshold 0.5 by default):

• mapper.map_cells(cell_names: list, threshold: float = 0.5, use_llm_selection: bool = False, llm_candidates_count: int = 10) → dict — for ad-hoc lookups (no AnnData).
• mapper.map_adata(adata, cell_name_col=None, threshold=0.5, new_col_name='cell_ontology') → dict — primary entry point; writes obs columns, returns the mapping result dict.
• mapper.map_cells_with_expansion(cell_names, threshold, expand_abbreviations=True, tissue_context, species, study_context, use_llm_selection=True, llm_candidates_count=10).
• mapper.map_adata_with_expansion(adata, cell_name_col=None, threshold=0.5, new_col_name='cell_ontology', expand_abbreviations=True, tissue_context=None, species=None, study_context=None, use_llm_selection=True, llm_candidates_count=10).
• mapper.map_cells_with_taxonomy(cell_names, threshold, expand_abbreviations=True, use_taxonomy=True, species, tissue_context, study_context, use_llm_selection=True, llm_candidates_count=10).
• mapper.map_adata_with_taxonomy(adata, cell_name_col=None, threshold=0.5, new_col_name='cell_ontology', expand_abbreviations=True, use_taxonomy=True, species=None, tissue_context=None, study_context=None).

Configuration & resources:

• mapper.setup_llm_expansion(api_type='openai'|'custom_openai', api_key=None, model='gpt-3.5-turbo', base_url=None, cache_file='abbreviation_cache.json', tissue_context=None, species='human', study_context=None, extra_params=None).
• mapper.load_cell_taxonomy_resource(taxonomy_file, species_filter=None) — adds Cell Taxonomy as a second ontology.
• mapper.load_ontology_mappings(popv_json_path) — load PopV-style CL ID mappings.
• mapper.create_ontology_resources(cl_obo_file, save_embeddings=True) — re-build the embeddings index.
• mapper.set_model(model_name, local_model_dir=None) — switch sentence-transformer.
• mapper.set_local_model(model_path) — load from a local model dir (offline / air-gapped).
• mapper.download_model() — manual model download.

Inspection:

• mapper.search_ontology_cells(keyword, case_sensitive=False, max_results=20) — keyword search over CL terms.
• mapper.list_ontology_cells(max_display=50, return_all=False) — list CL terms.
• mapper.browse_ontology_by_category(categories=None, max_per_category=10) — browse by anatomy / function category.
• mapper.find_similar_cells(cell_name, top_k=10) — top-k CL matches to a single name.
• mapper.find_similar_cells_taxonomy(cell_name, species=None, top_k=10) — same but in Cell Taxonomy.
• mapper.search_by_marker(markers, species=None, top_k=10) — find CT terms by marker overlap.
• mapper.get_cell_info(cell_name) / get_cell_info_taxonomy(cell_name, species=None) — full CL / CT entry.
• mapper.check_ontology_status() — diagnostic: ontology size, embedding state, model state.
• mapper.get_ontology_statistics(), mapper.get_statistics(mapping_results).
• mapper.print_mapping_summary(mapping_results, top_n=10), print_mapping_summary_taxonomy(...), print_mapping_summary_with_ids(...).
• mapper.show_expansion_summary(mapping_results) — what the LLM expansion did.
• mapper.test_abbreviation_detection(test_cases=None) — sanity-check abbreviation detection.

Persistence:

• mapper.save_embeddings(output_path=None) / load_embeddings(embeddings_path).
• mapper.save_mapping_results(mapping_results, output_file).
• mapper.clear_abbreviation_cache().

Boundary

Inside scope:

• Plain CL mapping via cosine similarity.
• LLM-driven abbreviation expansion before matching.
• Cell Taxonomy resource enhancement for species + tissue context.
• Diagnostic helpers (search, list, browse, similar).
• Persistence of embeddings + abbreviation cache.

Outside scope — separate skill:

• Building the cell-type annotation in the first place — single-cell-annotation skill (CellTypist / SCSA / gpt4celltype).
• LLM-only annotation (no CL mapping) — single-cell-annotation skill's gptcelltype branch.
• Cross-modality label transfer (RNA → ATAC) — see cross-modal-celltype-transfer skill.
• Reference-based label transfer (atlas → query) — see reference-label-transfer skill.
• MetaTiME tumour-microenvironment cell-state annotation — see single-cell-metatime-annotation.

Branch Selection

Plain vs expansion vs taxonomy

• Plain (map_adata): labels are full descriptive text (e.g. 'CD8+ effector memory T cell'). Don't pay the LLM cost.
• Expansion (map_adata_with_expansion): labels include abbreviations (e.g. 'TA-Early', 'TIL-1', 'IEL'). LLM expansion is critical for recall.
• Taxonomy (map_adata_with_taxonomy): cohort is species-specific and tissue-specific; you want both CL and species-resolved Cell Taxonomy matches. Pass species='Mus musculus', tissue_context='Gut'.

threshold (cosine similarity cutoff)

• 0.5 (default) — confidence floor; matches below this are reported as 'No clear match'.
• 0.3 — looser; useful with abbreviation expansion where similarity scores are systematically lower.
• 0.7 — stricter; only high-confidence matches.

use_llm_selection (in expansion mode)

• True (default in _with_expansion / _with_taxonomy): LLM picks the best CL match from the top-llm_candidates_count cosine candidates. Higher accuracy, costs LLM tokens.
• False: pure cosine; no LLM at the matching step.

llm_candidates_count

• 10 (default) — gives the LLM enough alternatives to choose from.
• 5 — cheaper LLM calls; less robust.
• 20+ — overkill; the LLM rarely benefits from >15 candidates.

api_type for setup_llm_expansion

• 'openai' — direct OpenAI API.
• 'custom_openai' — OpenAI-compatible endpoint (Azure, Ollama, OhMyGPT, vLLM-OpenAI). Pass base_url.

Sentence-transformer model

• 'all-mpnet-base-v2' (default) — best-quality general-purpose biomedical mapping at ~110M params.
• 'all-MiniLM-L6-v2' — much smaller (~22M); 5–10× faster but ~5 % lower mapping accuracy.
• Switch with mapper.set_model(model_name, local_model_dir) or pass to constructor.

Cache management

• embeddings_path='ontology_embeddings.pkl' — first call writes; subsequent calls read.
• LLM expansion cache: setup_llm_expansion(cache_file='abbreviation_cache.json'). Persistent across sessions; clear with mapper.clear_abbreviation_cache().
• Re-build embeddings when CL is updated: mapper.create_ontology_resources(cl_obo_file, save_embeddings=True).

Input Contract

• AnnData with obs[cell_name_col] populated; values are strings (cell-type labels). Empty / NaN entries skip mapping.
• For LLM expansion: outbound HTTPS access to OpenAI (or compatible); valid api_key. Costs ~$0.001 per cell-type label expanded; the cache makes repeated runs free.
• For Cell Taxonomy enhancement: download Cell_Taxonomy_resource.txt separately (URL in tutorial; ~50 MB) and pass to load_cell_taxonomy_resource.
• Sentence-transformer first run: ~110M-param model download (cached via HuggingFace; pass local_model_dir to control location).
• Hardware: CPU is sufficient (~30 s for ~3000-term embedding); GPU not required.

Minimal Execution Patterns

import omicverse as ov
import scanpy as sc

ov.plot_set()

# 1) Get the Cell Ontology
ov.single.download_cl(output_dir='new_ontology', filename='cl.json')

# 2) Construct mapper (first call computes embeddings; subsequent loads from cache)
mapper = ov.single.CellOntologyMapper(
    cl_obo_file='new_ontology/cl.json',
    embeddings_path='new_ontology/ontology_embeddings.pkl',
    local_model_dir='./my_models',
)

# 3) Plain mapping
mapping_results = mapper.map_adata(adata, cell_name_col='cell_label')
mapper.print_mapping_summary(mapping_results, top_n=15)

# 4) Inspect on UMAP
ov.pl.embedding(
    adata, basis='X_umap',
    color=['cell_label', 'cell_ontology', 'cell_ontology_cl_id'],
    wspace=0.55, ncols=2,
)

# LLM-expansion mode (handles abbreviations)
mapper.setup_llm_expansion(
    api_type='openai', model='gpt-4o-2024-11-20',
    api_key='sk-...',
    tissue_context='gut',
    species='mouse',
    study_context='Epithelial cells from small intestine and organoids; some Salmonella or H. polygyrus infected.',
)

# OR via a custom OpenAI-compatible endpoint
# mapper.setup_llm_expansion(
#     api_type='custom_openai',
#     api_key='sk-...',
#     model='gpt-4.1-2025-04-14',
#     base_url='https://api.ohmygpt.com/v1',
# )

mapping_results = mapper.map_adata_with_expansion(
    adata=adata,
    cell_name_col='cell_label',
    threshold=0.5,
    expand_abbreviations=True,
)
mapper.print_mapping_summary(mapping_results, top_n=15)
mapper.show_expansion_summary(mapping_results)

# Taxonomy-enhanced mode (species + tissue)
mapper.load_cell_taxonomy_resource(
    'new_ontology/Cell_Taxonomy_resource.txt',
    species_filter=['Homo sapiens', 'Mus musculus'],
)

enhanced_results = mapper.map_adata_with_taxonomy(
    adata,
    cell_name_col='cell_label',
    new_col_name='enhanced_cell_ontology',
    expand_abbreviations=True,
    use_taxonomy=True,
    species='Mus musculus',
    tissue_context='Gut',
    threshold=0.3,
)

ov.pl.embedding(
    adata, basis='X_umap',
    color=['cell_label', 'cell_ontology', 'enhanced_cell_ontology',
           'enhanced_cell_ontology_taxonomy_match',
           'enhanced_cell_ontology_ct_id'],
    wspace=0.55, ncols=2,
)

Validation

• After download_cl: cl.json exists at the requested path; size ~10 MB.
• After construction: mapper.check_ontology_status() reports the ontology size (~3000 CL terms) and embedding state (loaded vs to-build).
• After map_adata: obs['cell_ontology'] populated; (obs['cell_ontology_score'] >= threshold).sum() / len(obs) is the match rate. Healthy cohorts get >70 % at threshold=0.5. Lower rates suggest abbreviations — switch to expansion mode.
• After expansion mode: mapper.show_expansion_summary(...) tells you which labels got expanded; if zero labels were expanded, setup_llm_expansion wasn't called or the LLM rejected all.
• Cache freshness: re-running map_adata_with_expansion should hit the abbreviation cache for already-seen labels (no LLM cost). Watch the request count to confirm.
• Sanity check: pick a few labels manually and run mapper.find_similar_cells(<label>, top_k=10) — the top match should be biologically reasonable. If the top match is generic (e.g. 'native cell'), the label is too vague.

Resource Map

• See reference.md for compact copy-paste snippets per mode.
• See references/source-grounding.md for verified CellOntologyMapper constructor + the ~30 method signatures, plus the docstring backfill log (the class-level docstring was a single emoji line; expanded to full).
• For the upstream cell-type annotation that produces the input cell_label, see existing skills single-cell-annotation, reference-label-transfer, single-popv-annotation, omicverse-single-cell-metatime-annotation.
• For multi-annotator consensus (after CL mapping), see omicverse-single-cell-cellvote-consensus.

Examples

• "Map free-text cell labels in obs['cell_label'] to Cell Ontology terms with the default mpnet-base-v2 sentence transformer."
• "Cohort uses heavy abbreviations like 'TIL-1' / 'TA-Early' — set up LLM expansion with tissue_context='gut' / species='mouse' and re-map."
• "Add Cell Taxonomy enhancement for a mouse-gut cohort with threshold=0.3."
• "List the top-10 CL matches for a single label without modifying any AnnData (mapper.find_similar_cells('TIL-1', top_k=10))."

References

• Tutorial notebook: t_cellmatch.ipynb — full three-mode walkthrough on a mouse gut cohort.
• Cell Ontology: https://obofoundry.org/ontology/cl.html
• Cell Taxonomy resource: Jin et al. 2023.
• Live API verified — see references/source-grounding.md.